MAHALO: Unifying Offline Reinforcement Learning and Imitation Learning from Observations